Object detection in high-resolution microscopy is essential for automated bacterial analysis but demands computationally intensive models that challenge embedded deployment. This work addresses the problem of compressing large detectors for detecting Pseudomonas aeruginosa in crowded microscopic scenes, where very small and densely packed targets exacerbate the accuracy loss typically incurred by model compression. We explore guided feature knowledge distillation, focusing in particular on Masked Generative Distillation and Channel-Wise Distillation applied to the Yolo11 family, evaluating both direct teacher-student transfer and an assistant-teacher setup that bridges large capacity gaps. To better handle tiny objects, we introduce an overlapping \(640 \times 640\) pixel patching strategy that expands the effective training set. Experiments demonstrate that our lightweight Yolo11-N student improves from a baseline AP \(_{50\text {--}95}\) of 71.9 to 74.0 using a single assistant teacher, outperforming direct distillation results (73.3–73.5) while approaching the teacher performance of 75.5. These results establish that feature distillation combined with patch learning effectively compresses large microscopy detectors without sacrificing accuracy, yielding efficient models suitable for resource-constrained deployment.